Tokenizer

We trained our tokenizer using sentencepiece's unigram tokenizer. Then loaded the tokenizer as MT5TokenizerFast.

Model

We used MT5-base model.

Datasets

We used Code Search Net's dataset and some scrapped data from internet to train the model. We maintained a list of datasets where each dataset had codes of same language.

Plots

Train loss

Evaluation loss

Evaluation accuracy

Learning rate

Fine tuning (WIP)

We fine tuned the model with CodeXGLUE code-to-code-trans dataset, and scrapper data.